使用Keras Tuner优化CNN时获取pool超参数出现KeyError问题排查

问题现象 #

使用Keras Tuner优化CNN模型时，能正常获取滤波器、核大小的最优值，但获取池化大小（pool size）时触发如下错误：

File "C:\Users\...", line 180, in pp
    -optimal learning rate for the optimizer is {best_hps.get('learning_rate')}.""")

  File "C:\Users\anaconda3\envs\venv\lib\site-packages\keras_tuner\engine\hyperparameters\hyperparameters.py", line 241, in get
    raise KeyError(f"{name} does not exist.")

KeyError: 'pool does not exist.'

用户的实现代码：

def build_model(hp):
    model = Sequential()
    model.add(Conv2D(filters=hp.Int('conv_1_filter', min_value=32, max_value=128, step=16), 
                     kernel_size=hp.Choice('conv_1_kernel', values = [2, 3,5]),
                     strides=2, 
                     input_shape=(num_rows, num_columns, num_channels), 
                     activation='relu'))
    model.add(MaxPooling2D(pool_size=hp.Choice('pool', values = [2, 3])))

    model.add(Flatten())

    model.add(Dense(num_labels, activation='softmax'))

# Display model architecture summary 
    # model.summary()

# Compile the model
    model.compile(loss='categorical_crossentropy', metrics=['accuracy'], optimizer=keras.optimizers.Adam(hp.Choice('learning_rate', values=[1e-2, 1e-3])))
    return model

from kerastuner import RandomSearch
tuner = RandomSearch(build_model,
                  objective='val_accuracy',
                  max_trials = 5)
tuner.search(X_train, Y_train,epochs=3,validation_data=(X_train, Y_train),verbose = 1)

# Get the optimal hyperparameters
best_hps=tuner.get_best_hyperparameters(num_trials=1)[0]


print(f"""The hyperparameter search is complete. The parameters are as follow:
       -optimal filter size for the first layer is {best_hps.get('conv_1_filter')}
       -optimal pool size is {best_hps.get('pool')}
       -optimal kernel size for the first layer is {best_hps.get('conv_1_kernel')}
       -optimal learning rate for the optimizer is {best_hps.get('learning_rate')}.""")

解决方案 #

在RandomSearch初始化时添加overwrite=True参数，问题根源是程序复用了之前的旧调优结果，导致新添加的pool参数未被纳入搜索结果中。修改后的初始化代码如下：

tuner = RandomSearch(build_model,
                  objective='val_accuracy',
                  max_trials = 5,
                  overwrite=True)

内容的提问来源于stack exchange，提问作者A. Gehani